Assessing Scientific Theories: the Bayesian Approach

Assessing Scientific Theories: The Bayesian Approach∗ Radin Dardashtiy Stephan Hartmannz April 10, 2019 Abstract Scientists use a variety of methods to assess their theories. While experimental testing remains the gold standard, several other more controversial methods have been proposed, especially in fundamental physics. Amongst these methods are the use of analogue experiments and so-called non-empirical ways of theory-assessment such as the no- alternatives argument. But how can these methods themselves be assessed? Are they reliable guides to the truth, or are they of no help at all when it comes to assessing scientific theories? In this chapter, we de- velop a general Bayesian framework to scrutinize these new (as well as standard empirical) methods of assessing scientific theories and illustrate the proposed methodology by two detailed case studies. This allows us to explore under which conditions nontraditional ways of assessing scientific theories are successful and what can be done to improve them. ∗This is a penultimate version of Chapter 4 in: Dardashti, R., Dawid, R., and Thebault,´ K. (Eds.) (2019). Why Trust a Theory?: Epistemology of Fundamental Physics. Cambridge University Press. yInterdisziplinares¨ Zentrum fur¨ Wissenschafts- und Technikforschung/Philosophisches Seminar, Bergische Universitat¨ Wuppertal, 42119 Wuppertal (Germany) – https://radindardashti.weebly.com – [email protected]. zMunich Center for Mathematical Philosophy, LMU Munich, 80539 Munich (Germany) – http://www.stephanhartmann.org – [email protected]. 1 Contents 1 Introduction 2 2 Trusting Theories 4 3 Bayesian Confirmation Theory 6 4 Illustration 1: The No Alternatives Argument 9 5 Illustration 2: Analogue Experiments 12 6 Discussion 16 7 Conclusion 18 1 Introduction Scientific theories are used for a variety of purposes. For example, physi- cal theories such as classical mechanics and electrodynamics have important applications in engineering and technology, and we trust that this results in useful machines, stable bridges, and the like. Similarly, theories such as quan- tum mechanics and relativity theory have many applications as well. Beyond that, these theories provide us with an understanding of the world and address fundamental questions about space, time, and matter. Here we trust that the answers scientific theories give are reliable and that we have good reason to believe that the features of the world are similar to what the theories say about them. But why do we trust scientific theories, and what counts as evidence in favor of them? Clearly, the theories in question have to successfully relate to the outside world. But how, exactly, can they do this? The traditional answer to this question is that established scientific theories have been positively tested in experiments. In the simplest case, scientists derive a prediction from the theory under consideration, which is then found to obtain in a direct observation. Actual experimental tests are, of course, much more intricate, and the evalua- tion and interpretation of the data is a subtle and by no means trivial matter. One only needs to take a look at the experiments at the Large Hadron Col- lider (LHC) at the European Organization for Nuclear Research (CERN) near Geneva, and at the huge amount of workforce and data analysis involved there, 2 to realize how nontrivial actual experiments can be. However, there is no con- troversy among scientists and philosophers of science about the proposition that the conduct and analysis of experiments is the (only?) right way to assess theories. We trust theories because we trust the experimental procedures to test them. Philosophers of science have worked this idea out and formu- lated theories of confirmation (or corroboration) that make the corresponding intuition more precise. We discuss some of these proposals in the next two sections. While empirical testing is certainly important in science, some theories in fundamental science cannot (yet?) be tested empirically. String theory is a case in point. This arguably most fundamental scientific theory makes, so far, no empirically testable predictions, and even if it would make any distinct predictions, referring to predictions that are not also predictions of the theories it unifies (i.e., the Standard Model and General Relativity), these predictions could not be confirmed in a laboratory because the required energies are too high. The question, then, is what we should conclude from this. Are fundamental theories such as string theory not really scientific theories? Are they only mathematical theories? Many people would refrain from this conclusion and argue that string theory does, indeed, tell us something about our world. But why should we believe this? Are there other ways of assessing scientific theories, ways that go beyond empirical testing? Several well-known physicists think so (see, for example, Polchinski, this volume) and in a recent book, the philosopher Richard Dawid (2013) defends the view that there are non-empirical ways of assessing scientific theories. In his book, he gives a number of examples including the so-called No Alterna- tives Argument, which we discuss in detail in section 4. Another example is analogue experiments, examined in section 5. There is, however, no agree- ment among scientists regarding the viability of these nonstandard ways of assessing scientific theories. In a similar vein, these new methods do not fit into traditional philosophical accounts of confirmation (or corroboration) such as Hempel’s hypothetico-deductive (HD) model or Popper’s falsificationism. Interestingly, these deductivist accounts exclude indirect ways of assessing scientific theories from the beginning as, in those accounts, the evidence for a theory under consideration must be a deductive consequence of it and it must be observed. There is, however, an alternative philosophical framework available: Bayesian confirmation theory. We will show that it can be fruitfully employed to analyze potential cases of indirect confirmation such as the ones mentioned 3 previously. This will allow us to investigate under which conditions indirect confirmation works, and it will indicate which, if any, holes in a chain of rea- soning have to be closed if one wants to make a confirmation claim based on indirect evidence. By doing so, Bayesian confirmation theory helps the scien- tist (as well as the philosopher of science) better understand how, and under which conditions, fundamental scientific theories such as string theory can be assessed and, if successful, trusted. The remainder of this chapter is organized as follows. Section 2 consid- ers the traditional accounts of assessing scientific theories mentioned earlier and shows that they are inadequate to scrutinize indirect ways of assessing scientific theories. Section 3 introduces Bayesian confirmation theory and il- lustrates the basic mechanism of indirect confirmation. The following two sections present a detailed Bayesian account of two examples, namely the No Alternatives Argument (section 4) and analogue experiments (section 5). Sec- tion 6 provides a critical discussion of the vices and virtues of the Bayesian approach to indirect theory assessment. Finally, Section 7 concludes with a summary of our main results. 2 Trusting Theories Why do we trust a scientific theory? One reason for trusting a scientific theory is certainly that it accounts for all data available at the time in its domain of applicability. But this is not enough: We also expect a scientific theory to account for all future data in its domain of applicability. What, if anything, grounds the corresponding belief? What grounds the inference from the success of a theory for a finite set of data to the success of the theory for a larger set of data? After all, the future data are, by definition, not available yet, but we would nevertheless like to work with theories which we trust to be successful in the future. Recalling Hume’s problem of induction (Howson 2000), Popper argues that there is no ground for this belief at all. Such inferences are not justified, so all we can expect from science is that it identifies those theories that do not work. We can only falsify a proposed theory if it contradicts the available data. Pop- per’s theory is called falsificationism and it comes in a variety of versions. Ac- cording to naive falsificationism, a theory T is corroborated if an empirically testable prediction E of T (i.e., a deductive consequence of T) holds. Note that this is only a statement about the theory’s past performance, with no impli- cations for its future performance. If the empirically testable prediction does 4 not hold, then the theory is falsified and should be rejected and replaced by an alternative theory. As more sophisticated versions of falsificationism have the same main problem relevant for the present discussion as naive falsificationism, we do not have to discuss them here (see Pigliucci or Carroll, this volume). All that matters for us is the observation that, according to falsificationism, a theory can be corroborated only empirically. Hence, a falsificationist cannot make sense out of indirect ways of assessing scientific theories. These possible new ways of arguing for a scientific theory have to be dismissed from the beginning because they do not fit the proposed falsificationist methodology. We think that this is not a good reason to reject the new methods. It may well turn out that we come to the conclusion that none of them work, but the reason for this conclusion should not be that the new method does not fit our favorite account of theory assessment. Hempel, an inductivist, argued that we have grounds to believe that a well- confirmed theory can also be trusted in the future. Here is a concise summary of the main idea of the hypothetico-deductive (HD) model of confirmation that Hempel famously defended: General hypotheses in science as well as in everyday use are in- tended to enable us to anticipate future events; hence, it seems rea- sonable to count any prediction that is borne out by subsequent observation as confirming evidence for the hypothesis on which it is based, and any prediction that fails as disconfirming evidence.

Load more